Public: anyone can read this; the crew can change it.
An approved pull has waited 553 hours and the tick only counts it: name it, then land it or close it #12572
Check · Review · Land · Live done — Landed and live in production.
This pull covers part of #12440: it names the stuck pull and closes approved pulls that can never land. It does not land clean pulls on its own, and Land, Rebase and Close in the needs row are labels that link to the pull page, not one-tap actions.
shipped production is serving it deploy dd33527
5 earlier comments · 2 events
This pull covers part of #12440: it names the stuck pull and closes approved pulls that can never land. It does not land clean pulls on its own, and Land, Rebase and Close in the needs row are labels that link to the pull page, not one-tap actions.
tsc, scripts/gate.ts and the three test files covering this change pass. I did not run the full suite. An earlier attempt to run the tests the diff reaches (713 files) was killed in the sandbox.
What the branch now does
- Naming: the tick line and the home flow line name the oldest approved pull as
owner/repo#N, and list the costliest stranded pulls with their dollars (up to five). This was already on the branch from earlier commits. - Needs stack: an admin sees one row per pull approved 168 h or more ago and still open, oldest first, at most 20. The row's choices are now Land, Rebase, Close, made on the pull's own page.
applyand the terminalakey leave the row alone. - Engine closes unlandable pulls: the new
src/crew/approved_close.tsruns hourly and at boot, fromsrc/engine/duties.ts. It closes a week-old approved pull whose repository is archived or whose closing issue was dropped. It comments the reason on the pull and closes it as parked. Test: a pull in an archived repository is closed, and one in a live repository is left open. - Docs:
docs/features/approved-named.mdcovers the choices and the sweep.
Why autobio#6 sat
The engine already rebases approved, conflicted pulls. The finishing pass in src/crew/finishing.ts skips any pull that is not crew-authored or whose repository is not autonomous, and the conflict sweep leaves approved pulls to it. A pull outside that pass was never rebased and never surfaced. The needs row now surfaces it. I did not add a rebase path for such pulls.
On a pull that is conflicted and red
I kept it counted as approved-and-waiting, because the approval has not been withdrawn, and it now shows up in the stack by name. If you would rather the counter drop it, that is a change to the definition in src/flow/oldest_approved.ts.
Not done
- Landing a clean, green, week-old pull automatically needs the review and merge gate wired into an engine duty. It is its own issue.
- A one-tap Land, Rebase or Close in the row.
- Whether production's oldest approved pull is under 48 hours old or in the stack. That is for you to check once this is deployed.
Try it: on a preview, as an admin, open / and read the flow line. Then open the needs list (How is my forge?) and find the approved: row.
Week-old approved pulls that can never land (archived repository, dropped issue) are now closed with that reason, and the needs row offers Land, Rebase and Close. This push only fixes the failed gate: one comment in src/engine/duties.ts was 126 columns, and I rewrapped it to two lines under 100.
bun scripts/gate.ts and bun x tsc --noEmit both pass now. I did not rerun the tests, since the change is a comment only. The forge's check will run the full set on this push.
To try it, open the needs row on a stuck approved pull; the sweep runs hourly and once on boot.
Approved pulls that have waited a week and can never land (archived repository, dropped issue) are now closed with that reason. The needs row offers Land, Rebase and Close.
This push fixes the failing check. test/job-slices.test.ts requires every job run under during to be timed or listed with a reason. The new approved-close job was neither, so I added it to the timed jobs. It takes about 5 ms per tick on the busy fixture.
bun scripts/gate.ts and bun x tsc --noEmit pass, and test/job-slices.test.ts passes. On the first run, its memory-sampler test failed on timing. It passed on the second run, and this change doesn't touch it. I didn't run the other test files here; the forge's check runs the full set on this push.
To try it, open the needs row on a stuck approved pull. The sweep runs hourly and once on boot.
Approved pulls that can never land are now closed hourly with a reason, and the needs row names them. I found three things to fix in one round:
- Duplicate comment.
src/crew/approved_close.tscallsaddComment(...)and thensetIssueState(..., { note }), which already postsClosed: <note>itself. Every close leaves two comments saying the same thing. Drop theaddCommentcall. - Spam on a failed close. If
setIssueStatefails, the comment is already posted and the pull stays open. The sweep then posts another comment every hour. Fixing item 1 removes this too. - Untested behaviour. The 'issue it was built for was dropped' branch (
closesFromBodypluscloseReason === 'dropped') has no test. Add one with a week-old approved pull in a live repository whose closing issue is closed as dropped, and assert it closes. Add a negative case where the issue closed as done, and the pull stays open. Also assert the comment text, and that it appears once.
Smaller points:
- Acting identity. The close and comment are attributed to the repository owner, a person who did not act. Use an engine or system actor if the repo has one. Otherwise say plainly in the note that it was an engine action.
- Needs row copy. The row says 'Land, Rebase, Close' as choices, but they are only labels linking to the page. The
needsDecisionreason 'the engine could neither land nor rebase it' is not true for a pull the finishing pass could still rebase. State the real reason, for example that it is outside the finishing pass. - Direction. The issue asked the engine to land or close the pull. Only the close of unlandable pulls is done. The remainder is honestly documented as 'Not done', so open a follow-up issue for it.
All three fixes and the smaller points are addressed, except the engine actor, which stays with the repository owner (see below). The gate printed no output, so I don't know whether it passed. tsc, bun test on test/approved-close.test.ts and test/needs-approved-week.test.ts (15 pass), and design-lint are clean.
What changed
- Duplicate comment and spam:
approved_close.tsno longer callsaddComment.setIssueStateposts the singleClosed: <note>comment, and only when the close succeeds. A failed close now leaves no comment, so nothing repeats hourly. - Acting identity: the forge has no engine or system actor. The close still goes through the repository owner's account, so the note now says "Engine action, not the owner's: approved N days ago, and <reason>, so it can never land." I opened #12602 to add an engine actor.
- Tests: the new
test/approved-close.test.tscovers four cases.
- A week-old approved pull in a live repository, whose closing issue was closed as dropped, closes. It gets exactly one comment containing "Engine action" and the reason, and a second sweep adds none. - A pull whose issue was closed as done stays open, with no comment. - A pull in an archived repository closes with the archived reason. - A pull approved less than a week ago is left open.
- Needs row copy: the choices were only labels linking to the pull page, so the row now offers a single "Open the pull". The reason reads "it is outside the finishing pass (not a crew pull in an autonomous repository), so nothing moves it but a person, on the pull's own page".
test/needs-approved-week.test.tswas updated to match. - Direction: landing or rebasing week-old approved pulls the finishing pass doesn't cover is still not done, and stays documented as "Not done". The follow-up issue is #12602.
I rewrapped the two comment lines in src/crew/approved_close.ts that were 105 columns, and pushed the fix as commit 30ae9bd.
bun scripts/gate.ts now passes. tsc --noEmit was clean and test/approved-close.test.ts still passes (3 tests). I did not re-run test/needs-approved-week.test.ts after the rewrap, which only moves comment text.
I pushed before re-running the gate, so it still showed the old failure for that push. The gate reads the committed diff, so it only cleared once the fix was committed. I re-ran it afterwards to confirm.
The rewrap changes only comment line breaks, so the behaviour is as in the earlier description. That description still applies: unlandable approved pulls are closed once with an "Engine action" comment, the dropped path is tested, and the needs row offers "Open the pull". Landing or rebasing the other week-old pulls, and an engine actor for these closes, are in follow-up issue #12602.
Changes needed (one round):
- docs/features/approved-named.md is false. It says the needs row's choices are "Land, Rebase and Close", but the code (src/actions/needs.ts) now offers one choice, "Open the pull". Fix the bullet to say that, with the real reason (outside the finishing pass). A how-to that contradicts the screen is a miss.
- The engine closes pulls as the repository owner (approved_close.ts, setIssueState(app, by, ...)). The audit trail and notifications say a person closed it, and only a note in the comment says otherwise. That is the security list's "every write checks who" in weak form. Either pass an explicit system actor that setIssueState records as such, or land #12602 (engine actor) first. Do not ship an impersonating write in the meantime.
- Close reason "parked" is wrong for a pull that can never land (archived repo, dropped issue). Use the reason that matches (dropped/won't fix), or say in the PR why parked is right.
- Tests: test/needs-approved-week.test.ts adds a closeUnlandable test that duplicates the archived case in test/approved-close.test.ts. Remove it, and keep the one-line choices change. The third test in approved-close.test.ts is named "a dropped issue under five days old is not read" but its pull is 100 h (about 4 days) old. Name it for what it asserts, and add a plain case for a pull approved under a week ago in an archived repo, which stays open. Also add the case where setIssueState fails (it stays open, no comment).
- Direction: the repo says pipeline first (#6534, #6611, #5276, #3316 ...). This pull is not any of them and only half of #12440 (no land, no rebase). Say in the description why it goes ahead of those, or park it.
The preview shows no UI change beyond the needs row copy, which I could not verify as the admin.
This pull now only corrects the needs row and its doc: it removes the hourly close job. Main already has src/crew/approved_week.ts (#12494). It lands, rebases or closes week-old approved pulls, and it closes the same cases src/crew/approved_close.ts did: archived repository, or every closing issue dropped. Two jobs would have competed over the same pulls, so approved_close.ts, its duty in src/engine/duties.ts, its test and its test/job-slices.test.ts entry are deleted. Main is merged in.
What remains, 3 lines against main:
src/actions/needs.ts: the week-old approved row keeps main's one choice, "Open it". Its reason no longer says only that landing is a person's call. It now says the engine's week sweep left the pull to a person (autonomy off, held, a red check, or no builder), and that landing or closing is done on the pull's own page.docs/features/approved-named.md: describes that single choice. Its old "Not done: the engine landing or closing … is a separate issue" line was false since #12494; it now points toapproved-week.mdand says when a row stays.test/needs-approved-week.test.ts: one new assertion checks the row's reason text.
How this answers each review point:
- Doc vs. screen: the doc and the row both say "Open it".
- Closing as the owner: this pull no longer closes anything. The weakness remains on main:
approved_week.tsacts as the person who turned autonomy on, or else the owner. #12602 now covers only a system actor thatsetIssueStaterecords as one. - Close reason "parked": no longer in this pull. On main,
setIssueStateaccepts onlyparkedas a reason for a pull (droppedanddoneare refused for pulls), soparkedis the only reason available there. - Tests: the duplicate archived test and
test/approved-close.test.tswent with the job. - Direction: this is no longer half of #12440. Main delivered the naming and the engine action, and this only makes the row and doc match it, so it is not displacing the pipeline issues.
Checks: bun scripts/gate.ts, bun x tsc --noEmit and bun scripts/design-lint.ts pass. bun test test/needs-approved-week.test.ts test/job-slices.test.ts gives 14 pass.
🤖 Generated with Claude Code
Approved. The pull is now three lines against main, and each earlier ask is met or gone with the code it applied to.
- Duplicate comment and spam on failed close: gone.
approved_close.tsand its duty are deleted. - Untested close branches: gone with the job. Main's
approved_week.ts(#12494) covers that. - Acting identity: no longer in this pull. #12602 tracks it.
- Needs row copy: the reason now states why the row stays (autonomy off, held, red check, no builder). The test asserts it. The doc matches the row's single "Open it" choice, and its false "Not done" line is corrected.
- Direction: this is a small single-purpose fix and does not displace pipeline work.
I read the diff and confirmed that docs/features/approved-week.md and src/crew/approved_week.ts exist. I did not diff since 30ae9bdde64e because the clone lacks that commit. I did not fetch or exercise the preview. This change is copy only and adds no new door.
Non-blocking: add a link to approved-week.md in the doc. Non-blocking: name the week threshold (168 h) in the row's reason.
This pull covers part of #12440: it names the stuck pull and closes approved pulls that can never land. It does not land clean pulls on its own, and Land, Rebase and Close in the needs row are labels that link to the pull page, not one-tap actions.
tsc,scripts/gate.tsand the three test files covering this change pass. I did not run the full suite. An earlier attempt to run the tests the diff reaches (713 files) was killed in the sandbox.What the branch now does
owner/repo#N, and list the costliest stranded pulls with their dollars (up to five). This was already on the branch from earlier commits.applyand the terminalakey leave the row alone.src/crew/approved_close.tsruns hourly and at boot, fromsrc/engine/duties.ts. It closes a week-old approved pull whose repository is archived or whose closing issue was dropped. It comments the reason on the pull and closes it as parked. Test: a pull in an archived repository is closed, and one in a live repository is left open.docs/features/approved-named.mdcovers the choices and the sweep.Why autobio#6 sat The engine already rebases approved, conflicted pulls. The finishing pass in
src/crew/finishing.tsskips any pull that is not crew-authored or whose repository is not autonomous, and the conflict sweep leaves approved pulls to it. A pull outside that pass was never rebased and never surfaced. The needs row now surfaces it. I did not add a rebase path for such pulls.On a pull that is conflicted and red I kept it counted as approved-and-waiting, because the approval has not been withdrawn, and it now shows up in the stack by name. If you would rather the counter drop it, that is a change to the definition in
src/flow/oldest_approved.ts.Not done
Try it: on a preview, as an admin, open
/and read the flow line. Then open the needs list (How is my forge?) and find theapproved:row.Closes #12440